Your assistant builds in code now. This is where it lives.

Your assistant publishes over MCP, from the same conversation that wrote the thing. HTML and React pages serve as live, interactive pages; React compiles in the browser, so there is no build step and no deploy pipeline. Markdown, code, images, PDFs and Office documents sit beside them, each served the right way: a page as a page, a spreadsheet as a clean download. A file you already have drags in through an upload link, up to 200 MB, and a whole folder arrives as a folder.

Every publish of the same name is a new version underneath one unchanging address. The history is readable in the browser, each version with its date, its author and its note, and a writer can put any of them back. Nothing there deletes a version, so restoring is itself undoable. A save with no note gets one proposed from what actually changed.

A space is one client, one project or one team: its documents, their statuses, its own written context and its own discussion thread. The overview of a space is generated from real counts and real activity, never typed in, so a library of hundreds of pages still has a shape you can see. Six layouts to choose from: a file list, a dashboard, a showcase, a story, a handbook, a shelf.

Search reads the text inside documents, Office files included, and matches a document's owner and type as well as its title, saying which one matched. The Navigator puts one box over everything a member can reach: pages, documents, plans, cards, tables and the product's own help. Where AI is switched on, a workspace-wide question is answered from your own documents, through your own visibility, and the answer ends in places you can open.

The set that content governance actually needs, and the set a file server never had a column for. A review date is kept exactly as typed, so "end of Q3" survives and only a date that parses can become overdue. A document's state is a label other people rely on, so it gates nothing and says so. A check-out really does refuse a second writer, in one place, and an admin can always release it.

A comment can be pinned to a highlighted passage or an image, drawn over the page rather than written into it. Take to AI hands the document, its comments and its context back to Claude or ChatGPT, or to Claude Code and Codex with the repository it was written from already selected, so a review becomes the next draft without anybody re-explaining what the document was for.

A library that only grows is the pile again. The Winnow finds documents that restate one another by counted overlap and lets you keep one and mark the rest superseded; nothing is deleted. The Shelf-Mark proposes a folder, a type and a review date for whatever is sitting at the top level, using only words the workspace already uses. Patina shades the shelf by age, owner or type, so the stale corner is visible at a glance.

A dataset is a small table with a name and a shape. Bind it to a page and the page fetches it at one address every time it is opened, so the number on screen is the number in the table. The Live module lists every table, which pages each one feeds, who keeps it, and when it last changed.

The refresh lives in the assistant, not here. Set up a Routine that runs on Monday at eight, gathers the week's figures and updates the dataset, and the page is current before anyone is at their desk. There is nothing to host, nothing to deploy, and no credential to another system stored.

Open a table and its rows are editable in place: writers only, undo, and a stale editor refused rather than the last save winning in silence. It is deliberately not a spreadsheet. No formulas, no sorting, no column surgery, because the shape belongs to whoever created the table and the page that reads it depends on that shape holding.

Publish the customer's page from a template or from scratch, bind it to their dataset, and hand over one link. Next quarter the link is the same and the numbers are new. The client status page and weekly numbers templates ship this way: one ask, a live page, a dataset already bound and seeded.

Connect a plan to a dataset and the table is generated from the cards: statuses, owners and dates as they stand. A delivery dashboard for a customer then reads from the same plan the team runs, so it cannot drift from the truth, and a document can be connected to a card so the card says what it is tracking.

An embed puts the page inside the tool the team already lives in, in a strip that frames it with its name, its freshness and a way back. Paste the link anywhere and it unfurls with a card. It is outbound display only: nothing here signs in to the host system, and there is no sync to keep alive.

Pages arrange into a site with your branding, your accent and your theme, on your own subdomain, with paths made of names. Layouts cover a showcase for the polished thing, a story for the narrative, a handbook for the reference. The overview pages are generated from what is there, so a site never looks abandoned, and a second site exists only when you need a separate boundary.

A share link is a deliberate act. Label it, set an expiry, add a password so the accident of a wrong forward opens nothing, and add a handling notice in your own words ("Under NDA. Please do not pass this on."). Audience lists every link with its opens, and a revoke where the row is.

A portal scopes an audience to one space or a whole site. An open portal is a link; a named portal asks each visitor for a name and counts them separately, so a supplier can give each customer their own door into their own space and nothing beside it. Revoke one visitor, or the whole portal, from the same row.

See it as… sets your membership aside at the door and shows exactly what the outside sees, with every write refused while you look. It is the audit switch: no guessing what a client will find, no second browser, no incognito window.

Reach is the ledger of what left the workspace. Opens are counted per link and per page; visitors are never identified or followed across pages, and the product does not claim to know who opened a link. What it can say, it says plainly, and what it cannot, it does not dress up.

Pick the host platform from a list rather than typing an origin string, and the page is framed inside it with its name, its freshness and a way back. oEmbed and Open Graph make a pasted link unfurl anywhere with a card. The page stays one page, at one address, whoever is looking at it from where.

A plan is a cascade of cards as deep as the work needs: features, bugs, risks, decisions, ideas, each with its own word and icon. A card carries a note under an agreed shape, fields, a checklist, comments, links to what it waits on, and the documents the work produced. One call returns the whole thing to an assistant; another returns the project's direction, what was tried and abandoned, and which documents were written before the direction changed. That is the context your assistant lost, kept.

Take to AI on a card builds the prompt: the card, its documents, its checklist and the plan's folder, pointing at the working guide for a single task. Claude Code and ChatGPT Codex open with the repository the document was written from already selected. The guide tells the assistant to publish its artifacts back, attach them to the card, move the status honestly and record what it decided, so the next session starts from the card rather than from a blank conversation.

Describe a function, a role or an engagement and Profusia proposes the cards, with their rhythms, as a proposal rather than a plan. Aim the same at an existing card and it proposes one level of the cascade beneath it. Nothing becomes a row until a person accepts what is on the screen.

After a session with Claude Code, one sentence about what shipped becomes proposed status moves and checklist ticks on the cards that already exist, chosen by number from a list, so nothing invented can land. Review mode asks one question per card that has gone quiet: still on, moved, or closed. Rewind shows the plan as it stood on any date, replayed from its own history.

Group and sort the List, drag cards across the Board, drag bars on the Timeline where a date parses, read the cascade on the Map, and on Flow see what waits on what: a card whose blocker has not finished reads blocked wherever it appears, whatever its own status says. My work reads every card that names you, across every plan, with the past-due and blocked lanes beside it.

Every member surface is built to a phone contract: no sideways scroll, finger-sized controls, panes that take the whole screen. Open a card, take it to Claude Code, read what came back, move the status, pin the next one. The work is one thumb away, and the plan is the same plan on the desk the next morning.

Mint a link to a plan for someone outside; scope it to one lane so the sibling lane is nowhere in the page. Commenters' names and other plans' names are withheld. Before you send it, have the plan read as the recipient would and get a proofreader's three-part answer rather than a score. When a plan needs to leave, it exports as Markdown or CSV.